List your company
ResearchFormerly WizWand

Geometry Problem Solving on Geometry3K

88.8AccuracyMARS-GPS
24.21640.98357.7574.517Jun 8, 2025Jul 27, 2025Sep 15, 2025Nov 3, 2025Dec 23, 2025Feb 10, 2026Apr 1, 2026
Updated 6mo ago

Evaluation results

MethodLinks
MARS-GPSCategory=Ours2026.0488.8——
LANS (Diagram GT)Category=Neural Methods, trained on the larger PGPS9K dataset=true2026.0482.3——
GPT-5Decoding=Greedy (τ=0), Normalization=Exact-match, Resolution=224x224, Prompting strategy=Identical prompts2026.0279.03——
Pi-GPSCategory=Neural-symbolic Methods2026.0477.8——
Gemini-2.5-ProDecoding=Greedy (τ=0), Normalization=Exact-match, Resolution=224x224, Prompting strategy=Identical prompts2026.0277.04——
SCA-GPSCategory=Neural Methods2026.0476.7——
PGPSNet-v2-SCategory=Neural Methods, trained on the larger PGPS9K dataset=true2026.0476.4——
Claude 4.5 SonnetCategory=Proprietary LLMs2026.0475.8——
GeometryZero-VL-7BTraining=GCPO2025.0674.68——
Qwen2.5-VL-7B-Instruct + GRPOTraining=GRPO2025.0673.46——
GPT-5.2Category=Proprietary LLMs2026.0473.1——
Qwen2.5-VL-7B-Instruct + ToRLTraining=ToRL2025.0672.39——
Qwen2.5-VL-7B-Instruct + SFTTraining=SFT2025.0669.54——
Qwen3-VLModel detail=Qwen3-VL-32B-Thinking, Decoding=Greedy (τ=0), Normalization=Exact-match, Resolution=224x224, Prompting strategy=Identical prompts2026.0268.89——
GeoDRLCategory=Neural-symbolic Methods2026.0468.4——
E-GPSCategory=Neural-symbolic Methods2026.0467.9——
Qwen2.5-VL-7B-InstructBase Model=Qwen2.5-VL-7B-Instruct2025.0665.38——
GOLDCategory=Neural Methods, trained on the larger PGPS9K dataset=true2026.0462.7——
GPT-5Category=Proprietary LLMs2026.0461.5——
Gemini 2Category=MLLMs2026.0460.7——
GeoformerCategory=Neural Methods2026.0459.3——
NGSCategory=Neural Methods2026.0458.8——
GPT-4oCategory=MLLMs2026.0458.6——
Inter-GPSCategory=Neural-symbolic Methods2026.0457.5——
Claude 3.5 SonnetCategory=MLLMs2026.0456.4——
GRPOBase Model=Qwen2.5-VL-7B-Instruct2026.0355.4954.06—
GeometryZeroBase Model=Qwen2.5-VL-7B-Instruct2026.0353.7254.07—
A²POBase Model=Qwen2.5-VL-7B-Instruct2026.0353.6155.76—
ToRLBase Model=Qwen2.5-VL-7B-Instruct2026.0353.552.92—
A²POBase Model=Qwen2.5-VL-3B-Instruct2026.0353.0548.22—
GeometryZeroBase Model=Qwen2.5-VL-3B-Instruct2026.0352.7246.35—
VAPOModel Scale=7B2025.0951.3——
GRPOBase Model=Qwen2.5-VL-3B-Instruct2026.0350.7247.01—
MM-EurekaModel Scale=7B2025.0950.2——
PAPOModel Scale=7B, Decoding Strategy=greedy decoding2025.0949.7——
VL-RethinkerModel Scale=7B2025.0949.6——
ToRLBase Model=Qwen2.5-VL-3B-Instruct2026.0347.0644.71—
SFTBase Model=Qwen2.5-VL-7B-Instruct2026.0346.1749.02—
SFTBase Model=Qwen2.5-VL-3B-Instruct2026.0339.439.56—
Base ModelModel Scale=7B2025.0938.5——
Qwen-VLCategory=MLLMs2026.0426.7——
G-Llava-7BModel Size=7B2025.06——49.31
GeometryZero-1.5BModel Size=1.5B2025.06——60.24
GeometryZero-3BModel Size=3B2025.06——79.25
GeometryZero-7BModel Size=7B2025.06——78.81
GNS-Llava-1.5-7BModel Size=7B2025.06——62
Qwen2.5-1.5B-InstructModel Size=1.5B2025.06——41.76
Qwen2.5-1.5B-Instruct + GRPOModel Size=1.5B2025.06——53.35
Qwen2.5-1.5B-Instruct + SFTModel Size=1.5B2025.06——44.25
Qwen2.5-1.5B-Instruct + ToRLModel Size=1.5B2025.06——57.01
Qwen2.5-3B-InstructModel Size=3B2025.06——65.83
Qwen2.5-3B-Instruct + GRPOModel Size=3B2025.06——75.87
Qwen2.5-3B-Instruct + SFTModel Size=3B2025.06——71.65
Qwen2.5-3B-Instruct + ToRLModel Size=3B2025.06——76.31
Qwen2.5-7B-InstructModel Size=7B2025.06——70.99
Qwen2.5-7B-Instruct + GRPOModel Size=7B2025.06——79.03
Qwen2.5-7B-Instruct + SFTModel Size=7B2025.06——75.98
Qwen2.5-7B-Instruct + ToRLModel Size=7B2025.06——78.75